Check the article's claims before you share it
An article can contain accurate information with AI assistance, or false information written entirely by a person. If your decision depends on what it says, check the claims directly. An authorship estimate cannot do that work for you.
- Open the source behind an important claim. Check that the cited document or quotation supports what the article says.
- Check the date and the subject. An older report about a different company, place or event may not support the current headline.
- Look for a named author or accountable publisher, then check what reporting or first-hand material the article provides. A byline alone does not verify its contents.
- For a claim with consequences, look for corroboration that adds its own evidence. Several pages repeating the same unsupported statement do not resolve the question.
Using a scan on a longer article
You can paste one English article of at least 150 words into the free text scanner. Keep the text you intend to assess separate from navigation, comments and unrelated articles. Read any result alongside the article's sources and the publisher's account of how it was made.
The published results describe a mixed writing population and include results by length. We have not published a separate accuracy figure for every article category, and those length bands do not establish that news articles are our best-performing use case. A longer text can still receive a wrong or inconclusive result. The methodology explains how to interpret the score.
Why a list of AI words cannot settle authorship
Repeated phrases or an impersonal tone can make you want to investigate a piece. Those features also occur in human writing, and their absence does not establish that a person wrote it. Avoid accusing an author on the basis of one word or a familiar paragraph style. Ask about the sources, drafts or editorial process when those details matter.
Research such as Weber-Wulff and colleagues' 2023 evaluation documents limitations in the tools it tested. It is useful background on detection, not a current accuracy measurement for every product. Our own dated results are on the evidence page.
What to do with a short comment
A comment below our minimum cannot be scanned here. Do not combine unrelated comments to manufacture a longer document; a combined score cannot tell you who wrote the individual messages. Inspect the account's public history and compare any factual claim with its source. Repeated messages are something to investigate, not proof that the account is a bot.
A profile image offers a separate question about the file's origin. Our Content Credentials guide explains what verified credentials can show and why missing credentials do not establish authenticity. For product comments, the review guide gives checks relevant to a purchase.
Reading a result in context
A likely-AI result does not establish that an author intended to deceive you. A likely-human result does not establish that the article is true. Keep the evidence about its claims separate from the question of how its sentences were prepared. If two tools disagree, the guide to detector accuracy explains what to look for in their published evidence.
Sources
- Weber-Wulff et al., Testing of Detection Tools for AI-Generated Text (2023). Its findings apply to the tools and conditions evaluated in that study.
Common questions
Are the words on those lists, like delve and tapestry, proof of AI?
No single word was, and we would push back on that, since people used every word on those viral lists long before any of this. Where word choice started to mean something was when you added it up over a whole text, and that adding up is what a calibrated scan does (how the scoring works). So if someone pointed at one word and said that proved it, we would not give that much weight.
What about a single comment on social media?
A 15-word comment is below our input floor. Do not combine unrelated comments to force an authorship verdict. Check the account history and the claims themselves. A profile photo can be checked separately for Content Credentials, but missing credentials leave its origin unknown.
How much of the internet is AI-generated?
There is no single percentage that answers this for the whole internet. Measurements depend on the content sampled, the date and how AI involvement is defined. Bot-traffic statistics count web requests. They do not measure article authorship. Evaluate the particular piece and its sources instead.
Does an AI verdict mean the article is untrustworthy?
No, it would not, and we would push back on a reader who took it that way. What you had was an estimate of where the text came from, and it said nothing about how true the text was or how good it was. An article that was drafted by AI could be accurate, gone over by a person and useful, and one that a person wrote from start to finish could be wrong or dishonest. Help from AI is normal in publishing now, so for our part what we would want to know is who checked the facts, and if the byline stood behind them.